rosshemsley / gonn

A Neural Network library in Go

🧠 Gonn - Go Neural Network

Because there aren't enough half-baked neural network libraries in the world yet.

A bare-bones deep neural network written in Go. Contains just enough code to train a simple network and make predictions.

Uses gonum for linear algebra.

Example

// Load up a training set
x, err := mnist.LoadImagesGzipFile("data/train-images-idx3-ubyte.gz")
if err != nil {
    log.Fatalf("Failed to load images: %s", err)
}

y, err := mnist.LoadLabelsGzipFile("data/train-labels-idx1-ubyte.gz")
if err != nil {
    log.Fatalf("Failed to load labels: %s", err)
}

// Create a simple feed-forward network
_, cols := x.Dims()

dnn := nn.NewFeedForwardNetwork(
    nn.NewFullyConnectedLayer(cols, 10),
    nn.NewSoftMaxLayer(),
)

// Train using stochastic gradient descent.
sgd.SGD(x, y, nn.L2Loss, dnn, sgd.WithBatchSize(256), sgd.WithEpochs(10))

Examples

This project uses go modules. If you have go1.11 or above, you can try this out by running

go get github.com/rosshemsley/gonn

If you have your go bin in your bath, you can now try the examples with

$ gonn <example name>

⚠️ Warning: this code is very much a toy implementation at the moment. You probably shouldn't be trying to use it.